From generated molecule to synthesis route, in one workflow

Vecura now runs Selenium, the retrosynthesis model from b12 Labs, as the final step of its discovery workflow, so a molecule you generate comes back with routes for making it. Here is what that looked like on a real run against NEK2.

Why the route matters

Retrosynthesis is how chemists plan a synthesis backwards. You start from the molecule you want, cut it apart at the bonds you know how to form, and keep going until every piece is something you can buy. Done well, it tells you early on whether a molecule can be made at all and where the difficult steps are going to be, before anyone has spent time or money in the lab.

That question comes up all the time in a computational discovery program. A generative model can propose a molecule with a lovely binding pose, but the pose is worth very little if nobody can make the compound at a sensible cost with the chemistry that is available. For a long time this check was done by experienced chemists working through named reactions by hand, which is slow and expensive, and in practice it rarely gets past the first one or two routes that look plausible. This is the gap that a retrosynthesis model like Selenium is built to close.

What Selenium does

Selenium is the retrosynthesis model built by b12 Labs. Instead of applying a fixed set of reaction rules, it works as an agent that reasons over known reactions, checks which starting materials can actually be bought and weighs what each option would cost. For a single target it returns several routes rather than one, ranked on step count, expected cost, likely yields and how easy the intermediates are to source, so a chemist can pick the route that fits the equipment, expertise and suppliers they have. It also handles stereochemistry properly, which matters for pharmaceutical compounds where a single stereocenter can decide whether a molecule is active or not.

Vecura now calls Selenium directly from its workflow. You can go from choosing a target through molecule generation and ADMET screening to a set of synthesis routes in one pipeline, without leaving the platform. Here is how that played out on a real example against NEK2 (Fig. 1), and you can run the same workflow on your own target here.

Figure 1. The five-stage workflow, where target resolution, structure retrieval, molecule generation and ADMET screening run in Vecura and the final retrosynthesis step calls Selenium from b12 Labs.

The run

We targeted NEK2, a serine/threonine kinase that regulates centrosome separation at the start of mitosis and is overexpressed in several cancers.

The workflow resolved the target to UniProt P51955 and pulled an experimental structure, PDB 2W5A, rather than falling back on a predicted model. That matters for everything that follows, because de novo generation is only as good as the pocket it is seeded on, and a co-crystal structure gives a far more faithful picture of the pocket than a prediction does.

PocketXMol then generated ten molecules, seeded on the 2W5A reference ligand within a 20 Å radius. The set was structurally diverse and included aminopyrimidines, an indole biaryl amine, alkynyl heterocycles, aryl amides, cyanopyridines and an unsaturated lactam. Several of those chemotypes line up with known ATP-competitive NEK2 inhibitors, which suggests the generator stayed anchored to real features of the pocket instead of wandering off into novelty for its own sake.

What got through

Nine of the ten structures produced valid SMILES, and those nine went through ADMET screening in two stages. All nine passed the drug-likeness stage, which covers molecular weight, logP, hydrogen bonding, Lipinski compliance and QED. The toxicity panel was a much higher bar, because a compound had to clear hERG, AMES, DILI, ClinTox and LD50 all at once.

Only one compound managed that, and the predicted liver injury model did most of the filtering on its own, which is a finding that deserves its own post at some point. For this piece, what matters is that a single molecule came out the other side, and that raises the obvious question of whether anyone can actually make it.

The handoff

The molecule that survived was a chiral unsaturated δ-lactam with a methoxyphenyl substituent, an exocyclic vinyl carbinol and one stereocenter. It is not a trivial thing to make, and it is exactly the kind of structure that a generative model proposes without a second thought and a chemist looks at with a raised eyebrow.

Vecura handed it straight to Selenium, with no file export, no reformatting and no separate login, and four routes came back. Each one is ranked, broken down into steps and comes with a feasibility assessment attached.

Table 1. Retrosynthetic routes returned by Selenium (b12 Labs) for the selected compound, ranked by score.

 

Why the ranking is the useful part

A list of possible routes is interesting to look at, but a ranked list with a feasibility call attached to each route is something you can act on, and this table shows the difference.

The top route builds the lactam by ring-closing metathesis, which is well precedented for six-membered nitrogen heterocycles of this kind, and the diene and amide precursors it needs come from standard building blocks. At eight steps and rated likely feasible, it is a route you can cost, schedule and hand to a chemist.

The fourth route is the shortest in the set at seven steps, and it is flagged as high risk. A chemist can see why straight away, because forming an oxime and running a Beckmann rearrangement next to a vinyl carbinol is asking for chemo selectivity problems. Anyone choosing routes on step count alone would have picked this one and found out about the problem in the lab.

The second route came back without a feasibility annotation, and that is useful in its own way, because it tells you to have someone review the route by hand before committing resources, rather than guessing in either direction.

The whole retrosynthesis step ran unattended and took about three and a half hours.

Why we integrated rather than built

Retrosynthesis is a hard, specialized problem with decades of prior work behind it, and only a handful of teams do it really well. b12 Labs came out of EPFL's Laboratory of Artificial Chemical Intelligence, and its platform exists to turn ambitious molecules into practical routes from starting materials you can buy.

We think a discovery platform should call the best available tool at each step rather than build a weaker version of everything itself. Vecura handles target resolution, structure retrieval, generation, property filtering and the reasoning that ties those steps together, and for synthesis planning we call b12 Labs and get back something better than we would have built on our own.

For you, that means a workflow where a generated molecule arrives with its route already attached, so the question of whether anyone can make it gets answered in the same session in which it was designed.

Scope. This run was computational from start to finish. No binding affinities or potency values were generated, so nothing here shows that these compounds engage NEK2, and that is the next experiment rather than a conclusion from this one. The synthesis routes are predictions and should be reviewed by a chemist before any material is ordered.

 

Bring your target to Vecura and get generated chemistry with synthesis routes attached, in one workflow.